Abstract
Nowadays PC-cluster architectures are widely accepted forparallel computing.In a PC-cluster system, memories are physicallydistributed.To harness thecomputational power of a distributed-memoryPC-cluster, a user must writeefficient software for the machine by hand. The absence of globaladdress space makes the process laborious because theuser must manually assign computations to processors, distribute data acrossprocessors andexplicitly manage communication among processors. This paperintroduces ourrecent three related studies on parallel computing. The first isem a V-Busbased PC-cluster in which all PCs are interconnected through V-Busnetworkcards. The second is the implementation of an one-sidedcommunicationlibrary on our PC-cluster system, which provides the user with aview ofglobal address space on top of the distributed-memory PC-cluster.Thisencapsulation of global address allows the user to write shared memory codeonour cluster system, which simplifies programming our cluster because theuserdoes not need to explicitly cope with distributed memories. The third is aparallelizing compiler that automatically translates a sequential code toashared-memory code for the PC-cluster. This compiler not only enables legacycodeto be compiled for our cluster, but also further simplifies programmingourcluster by allowing the user to continue using conventional sequentiallanguagessuch as Fortran 77.In this work, the compiler was optimizedparticularlyfor our V-Bus based PC-cluster. The paper also reports ourexperimental resultswith the compiler on our PC-cluster system.